26 #if defined(DATA_TYPE) && defined(ELEMENT_SIZE) 29 #define COND_DATA_TYPE char 30 #elif ELEMENT_SIZE == 2 31 #define COND_DATA_TYPE short 32 #elif ELEMENT_SIZE == 4 33 #define COND_DATA_TYPE int 35 #error "Element size not support" 36 #endif // ELEMENT_SIZE 38 #if defined(CONVOLVED_WIDTH) && defined(STRIDE_Y) && defined(SRC_DEPTH) 67 __kernel
void im2col1x1_stridex1_nchw(
69 #
if defined(NUM_GROUPS)
77 const uint xc = get_global_id(0) * 4;
78 const uint yc = get_global_id(1);
79 const uint ch = get_global_id(2) % SRC_DEPTH;
80 const uint batch = get_global_id(2) / SRC_DEPTH;
84 uint4 xc_clamped = xc + (uint4)(0, 1, 2, 3);
89 xc_clamped =
select((uint4)xc, xc_clamped, convert_int4(cond0));
93 const uint yi = yc * STRIDE_Y;
97 #if defined(NUM_GROUPS) 98 const uint xo = ch % (SRC_DEPTH / NUM_GROUPS);
99 const uint zo = ch / (SRC_DEPTH / NUM_GROUPS);
100 #else // defined(NUM_GROUPS) 102 #endif // defined(NUM_GROUPS) 103 const uint4 yo = xc_clamped + yc * CONVOLVED_WIDTH;
106 __global uchar *input_ptr = src_ptr + src_offset_first_element_in_bytes + xi * src_stride_x + yi * src_stride_y + ch * src_stride_z + batch * src_stride_w;
107 #if defined(NUM_GROUPS) 108 __global uchar *output_ptr = dst_ptr + dst_offset_first_element_in_bytes + xo * dst_stride_x + zo * dst_stride_z + batch * dst_stride_w;
109 #else // defined(NUM_GROUPS) 110 __global uchar *output_ptr = dst_ptr + dst_offset_first_element_in_bytes + xo * dst_stride_x + batch * dst_stride_w;
111 #endif // defined(NUM_GROUPS) 114 data = vload4(0, (__global
DATA_TYPE *)input_ptr);
119 *(__global DATA_TYPE *)(output_ptr + yo.s0 * dst_stride_y) = data.s0;
120 *(__global DATA_TYPE *)(output_ptr + yo.s1 * dst_stride_y) = data.s1;
121 *(__global DATA_TYPE *)(output_ptr + yo.s2 * dst_stride_y) = data.s2;
122 *(__global DATA_TYPE *)(output_ptr + yo.s3 * dst_stride_y) = data.s3;
125 #if defined(NUM_GROUPS) 126 if(xo == (SRC_DEPTH / NUM_GROUPS - 1))
127 #else // defined(NUM_GROUPS) 128 if(ch == (SRC_DEPTH - 1))
129 #endif // defined(NUM_GROUPS) 131 *((__global DATA_TYPE *)(output_ptr + yo.s0 * dst_stride_y) + 1) = 1.0f;
132 *((__global DATA_TYPE *)(output_ptr + yo.s1 * dst_stride_y) + 1) = 1.0f;
133 *((__global DATA_TYPE *)(output_ptr + yo.s2 * dst_stride_y) + 1) = 1.0f;
134 *((__global DATA_TYPE *)(output_ptr + yo.s3 * dst_stride_y) + 1) = 1.0f;
138 #endif // defined(CONVOLVED_WIDTH) && defined(STRIDE_Y) && defined(SRC_DEPTH) 140 #if defined(CONVOLVED_WIDTH) && defined(SRC_WIDTH) && defined(SRC_HEIGHT) && defined(STRIDE_X) && defined(STRIDE_Y) && defined(SRC_DEPTH) && defined(PAD_LEFT) && defined(PAD_RIGHT) && defined(PAD_TOP) && defined(PAD_BOTTOM) && defined(PAD_VALUE) 141 #if defined(DILATION_X) && defined(DILATION_Y) 174 __kernel
void im2col_generic_nchw(
176 #
if defined(NUM_GROUPS)
184 const int xc = get_global_id(0);
185 const int yc = get_global_id(1);
186 const int ch = get_global_id(2) % SRC_DEPTH;
187 const int batch = get_global_id(2) / SRC_DEPTH;
190 const int xi = xc * STRIDE_X - PAD_LEFT;
191 const int yi = yc * STRIDE_Y - PAD_TOP;
194 #if defined(NUM_GROUPS) 195 const int xo = (ch % (SRC_DEPTH / NUM_GROUPS)) * KERNEL_WIDTH * KERNEL_HEIGHT;
196 const int zo = ch / (SRC_DEPTH / NUM_GROUPS);
197 #else // defined(NUM_GROUPS) 198 const int xo = ch * KERNEL_WIDTH * KERNEL_HEIGHT;
199 #endif // defined(NUM_GROUPS) 200 const int yo = xc + yc * CONVOLVED_WIDTH;
202 __global uchar *input_ptr = src_ptr + src_offset_first_element_in_bytes + ch * src_stride_z + batch * src_stride_w;
203 #if defined(NUM_GROUPS) 204 __global DATA_TYPE *output_ptr = ((__global DATA_TYPE *)(dst_ptr + dst_offset_first_element_in_bytes + yo * dst_stride_y + zo * dst_stride_z + batch * dst_stride_w)) + xo;
205 #else // defined(NUM_GROUPS) 206 __global DATA_TYPE *output_ptr = ((__global DATA_TYPE *)(dst_ptr + dst_offset_first_element_in_bytes + yo * dst_stride_y + batch * dst_stride_w)) + xo;
207 #endif // defined(NUM_GROUPS) 210 for(
int yk = 0; yk < KERNEL_HEIGHT; ++yk)
212 int y = yi + yk * DILATION_Y;
213 for(
int xk = 0; xk < KERNEL_WIDTH; ++xk, ++output_ptr)
215 int x = xi + xk * DILATION_X;
216 #if PAD_LEFT == 0 && PAD_TOP == 0 && PAD_RIGHT == 0 && PAD_BOTTOM == 0 217 *output_ptr = *((__global DATA_TYPE *)(input_ptr + x * src_stride_x + y * src_stride_y));
218 #else // PAD_LEFT == 0 && PAD_TOP == 0 && PAD_RIGHT == 0 && PAD_BOTTOM == 0 219 if(x < 0 || x >= SRC_WIDTH || y < 0 || y >= SRC_HEIGHT)
221 *output_ptr = PAD_VALUE;
225 *output_ptr = *((__global DATA_TYPE *)(input_ptr + x * src_stride_x + y * src_stride_y));
227 #endif // PAD_LEFT == 0 && PAD_TOP == 0 && PAD_RIGHT == 0 && PAD_BOTTOM == 0 232 #if defined(NUM_GROUPS) 233 if((xo / (KERNEL_WIDTH * KERNEL_HEIGHT)) == (SRC_DEPTH / NUM_GROUPS - 1))
235 if(ch == (SRC_DEPTH - 1))
242 #endif // defined(DILATION_X) && defined(DILATION_Y) 274 __kernel
void im2col3x3_nchw(
276 #
if defined(NUM_GROUPS)
284 const int xc = get_global_id(0);
285 const int yc = get_global_id(1);
286 const int ch = get_global_id(2) % SRC_DEPTH;
287 const int batch = get_global_id(2) / SRC_DEPTH;
290 const int xi = xc * STRIDE_X - PAD_LEFT;
291 const int yi = yc * STRIDE_Y - PAD_TOP;
294 #if defined(NUM_GROUPS) 295 const int xo = (ch % (SRC_DEPTH / NUM_GROUPS)) * 9;
296 const int zo = ch / (SRC_DEPTH / NUM_GROUPS);
297 #else // defined(NUM_GROUPS) 298 const int xo = ch * 9;
299 #endif // defined(NUM_GROUPS) 300 const int yo = xc + yc * CONVOLVED_WIDTH;
303 __global uchar *input_ptr = src_ptr + src_offset_first_element_in_bytes + xi * (int)src_stride_x + yi * (
int)src_stride_y + ch * src_stride_z + batch * src_stride_w;
304 #if defined(NUM_GROUPS) 305 __global uchar *output_ptr = dst_ptr + dst_offset_first_element_in_bytes + xo * dst_stride_x + yo * dst_stride_y + zo * dst_stride_z + batch * dst_stride_w;
306 #else // defined(NUM_GROUPS) 307 __global uchar *output_ptr = dst_ptr + dst_offset_first_element_in_bytes + xo * dst_stride_x + yo * dst_stride_y + batch * dst_stride_w;
308 #endif // defined(NUM_GROUPS) 311 row0 = vload3(0, (__global DATA_TYPE *)(input_ptr + 0 * src_stride_y));
313 row1 = vload3(0, (__global DATA_TYPE *)(input_ptr + 1 * src_stride_y));
315 row2 = vload3(0, (__global DATA_TYPE *)(input_ptr + 2 * src_stride_y));
317 #if PAD_LEFT != 0 || PAD_TOP != 0 || PAD_RIGHT != 0 || PAD_BOTTOM != 0 319 int3 x = (int3)xi + (int3)(0, 1, 2);
320 int3 y = (int3)yi + (int3)(0, 1, 2);
323 cond0 =
CONVERT((x >= (int3)0 && x < (int3)SRC_WIDTH && (int3)(y.s0 >= 0 && y.s0 < SRC_HEIGHT)),
VEC_DATA_TYPE(COND_DATA_TYPE, 3));
325 cond1 =
CONVERT((x >= (int3)0 && x < (int3)SRC_WIDTH && (int3)(y.s1 >= 0 && y.s1 < SRC_HEIGHT)),
VEC_DATA_TYPE(COND_DATA_TYPE, 3));
327 cond2 =
CONVERT((x >= (int3)0 && x < (int3)SRC_WIDTH && (int3)(y.s2 >= 0 && y.s2 < SRC_HEIGHT)),
VEC_DATA_TYPE(COND_DATA_TYPE, 3));
332 #endif // PAD_LEFT != 0 || PAD_TOP != 0 || PAD_RIGHT != 0 || PAD_BOTTOM != 0 334 vstore8((
VEC_DATA_TYPE(DATA_TYPE, 8))(row0.s012, row1.s012, row2.s01), 0, (__global DATA_TYPE *)output_ptr);
335 *((__global DATA_TYPE *)output_ptr + 8) = row2.s2;
338 #if defined(NUM_GROUPS) 339 if((xo / 9) == (SRC_DEPTH / NUM_GROUPS - 1))
340 #else // defined(NUM_GROUPS) 341 if(ch == (SRC_DEPTH - 1))
344 *((__global DATA_TYPE *)output_ptr + 9) = 1.0f;
380 __kernel
void im2col5x5_nchw(
382 #
if defined(NUM_GROUPS)
390 const int xc = get_global_id(0);
391 const int yc = get_global_id(1);
392 const int ch = get_global_id(2) % SRC_DEPTH;
393 const int batch = get_global_id(2) / SRC_DEPTH;
396 const int xi = xc * STRIDE_X - PAD_LEFT;
397 const int yi = yc * STRIDE_Y - PAD_TOP;
400 #if defined(NUM_GROUPS) 401 const int xo = (ch % (SRC_DEPTH / NUM_GROUPS)) * 25;
402 const int zo = ch / (SRC_DEPTH / NUM_GROUPS);
403 #else // defined(NUM_GROUPS) 404 const int xo = ch * 25;
405 #endif // defined(NUM_GROUPS) 406 const int yo = xc + yc * CONVOLVED_WIDTH;
408 #if PAD_LEFT != 0 || PAD_TOP != 0 || PAD_RIGHT != 0 || PAD_BOTTOM != 0 410 int4 x0 = (int4)xi + (int4)(0, 1, 2, 3);
411 int4 y0 = (int4)yi + (int4)(0, 1, 2, 3);
417 x0_condition =
CONVERT((x0 >= (int4)0 && x0 < (int4)SRC_WIDTH),
VEC_DATA_TYPE(COND_DATA_TYPE, 4));
419 y0_condition =
CONVERT((y0 >= (int4)0 && y0 < (int4)SRC_HEIGHT),
VEC_DATA_TYPE(COND_DATA_TYPE, 4));
420 COND_DATA_TYPE x1_condition = (COND_DATA_TYPE)(x1 >= 0 && x1 < SRC_WIDTH);
421 COND_DATA_TYPE y1_condition = (COND_DATA_TYPE)(y1 >= 0 && y1 < SRC_HEIGHT);
422 #endif // PAD_LEFT != 0 || PAD_TOP != 0 || PAD_RIGHT != 0 || PAD_BOTTOM != 0 425 __global uchar *input_ptr = src_ptr + src_offset_first_element_in_bytes + xi * (int)src_stride_x + yi * (
int)src_stride_y + ch * src_stride_z + batch * src_stride_w;
426 #if defined(NUM_GROUPS) 427 __global uchar *output_ptr = dst_ptr + dst_offset_first_element_in_bytes + xo * dst_stride_x + yo * dst_stride_y + zo * dst_stride_z + batch * dst_stride_w;
428 #else // defined(NUM_GROUPS) 429 __global uchar *output_ptr = dst_ptr + dst_offset_first_element_in_bytes + xo * dst_stride_x + yo * dst_stride_y + batch * dst_stride_w;
430 #endif // defined(NUM_GROUPS) 434 row00 = vload4(0, (__global DATA_TYPE *)input_ptr);
436 row01 = *((__global DATA_TYPE *)input_ptr + 4);
438 input_ptr += src_stride_y;
441 row10 = vload4(0, (__global DATA_TYPE *)input_ptr);
443 row11 = *((__global DATA_TYPE *)input_ptr + 4);
445 #if PAD_LEFT != 0 || PAD_TOP != 0 || PAD_RIGHT != 0 || PAD_BOTTOM != 0 447 cond00 = x0_condition && (
VEC_DATA_TYPE(COND_DATA_TYPE, 4))y0_condition.s0;
449 cond10 = x0_condition && (
VEC_DATA_TYPE(COND_DATA_TYPE, 4))y0_condition.s1;
450 COND_DATA_TYPE cond01 = (COND_DATA_TYPE)(x1_condition && y0_condition.s0);
451 COND_DATA_TYPE cond11 = (COND_DATA_TYPE)(x1_condition && y0_condition.s1);
456 row01 =
select((DATA_TYPE)PAD_VALUE, row01, cond01);
457 row11 =
select((DATA_TYPE)PAD_VALUE, row11, cond11);
458 #endif // PAD_LEFT != 0 || PAD_TOP != 0 || PAD_RIGHT != 0 || PAD_BOTTOM != 0 462 0, (__global DATA_TYPE *)output_ptr);
463 vstore2((
VEC_DATA_TYPE(DATA_TYPE, 2))(row10.s3, row11), 0, (__global DATA_TYPE *)output_ptr + 8);
465 input_ptr += src_stride_y;
466 output_ptr += 10 * dst_stride_x;
471 row00 = vload4(0, (__global DATA_TYPE *)input_ptr);
473 row01 = *((__global DATA_TYPE *)input_ptr + 4);
475 input_ptr += src_stride_y;
478 row10 = vload4(0, (__global DATA_TYPE *)input_ptr);
480 row11 = *((__global DATA_TYPE *)input_ptr + 4);
482 #if PAD_LEFT != 0 || PAD_TOP != 0 || PAD_RIGHT != 0 || PAD_BOTTOM != 0 484 cond00 = x0_condition && (
VEC_DATA_TYPE(COND_DATA_TYPE, 4))y0_condition.s2;
486 cond10 = x0_condition && (
VEC_DATA_TYPE(COND_DATA_TYPE, 4))y0_condition.s3;
487 COND_DATA_TYPE cond01 = (COND_DATA_TYPE)(x1_condition && y0_condition.s2);
488 COND_DATA_TYPE cond11 = (COND_DATA_TYPE)(x1_condition && y0_condition.s3);
493 row01 =
select((DATA_TYPE)PAD_VALUE, row01, cond01);
494 row11 =
select((DATA_TYPE)PAD_VALUE, row11, cond11);
495 #endif // PAD_LEFT != 0 || PAD_TOP != 0 || PAD_RIGHT != 0 || PAD_BOTTOM != 0 499 0, (__global DATA_TYPE *)output_ptr);
500 vstore2((
VEC_DATA_TYPE(DATA_TYPE, 2))(row10.s3, row11), 0, (__global DATA_TYPE *)output_ptr + 8);
502 input_ptr += src_stride_y;
503 output_ptr += 10 * dst_stride_x;
508 row00 = vload4(0, (__global DATA_TYPE *)input_ptr);
510 row01 = *((__global DATA_TYPE *)input_ptr + 4);
512 input_ptr += src_stride_y;
514 #if PAD_LEFT != 0 || PAD_TOP != 0 || PAD_RIGHT != 0 || PAD_BOTTOM != 0 516 cond00 = x0_condition && (
VEC_DATA_TYPE(COND_DATA_TYPE, 4))y1_condition;
517 COND_DATA_TYPE cond01 = (COND_DATA_TYPE)(x1_condition && y1_condition);
521 row01 =
select((DATA_TYPE)PAD_VALUE, row01, cond01);
522 #endif // PAD_LEFT != 0 || PAD_TOP != 0 || PAD_RIGHT != 0 || PAD_BOTTOM != 0 524 vstore4(row00, 0, (__global DATA_TYPE *)output_ptr);
525 *((__global DATA_TYPE *)output_ptr + 4) = row01;
527 output_ptr += 5 * dst_stride_x;
531 #if defined(NUM_GROUPS) 532 if((xo / 25) == (SRC_DEPTH / NUM_GROUPS - 1))
533 #else // defined(NUM_GROUPS) 534 if(ch == (SRC_DEPTH - 1))
537 *((__global DATA_TYPE *)output_ptr) = 1.0f;
541 #endif // defined(CONVOLVED_WIDTH) && defined(SRC_WIDTH) && defined(SRC_HEIGHT) && defined(STRIDE_X) && defined(STRIDE_Y) && defined(SRC_DEPTH) && defined(PAD_LEFT) && defined(PAD_RIGHT) && defined(PAD_TOP) && defined(PAD_BOTTOM) && defined(PAD_VALUE) 543 #if defined(CONVOLVED_WIDTH) && defined(STRIDE_X) && defined(STRIDE_Y) && defined(SRC_DEPTH) 572 __kernel
void im2col11x11_padx0_pady0_nchw(
574 #
if defined(NUM_GROUPS)
582 const int xc = get_global_id(0);
583 const int yc = get_global_id(1);
584 const int ch = get_global_id(2) % SRC_DEPTH;
585 const int batch = get_global_id(2) / SRC_DEPTH;
588 const int xi = xc * STRIDE_X;
589 const int yi = yc * STRIDE_Y;
592 #if defined(NUM_GROUPS) 593 const int xo = (ch % (SRC_DEPTH / NUM_GROUPS)) * 121;
594 const int zo = ch / (SRC_DEPTH / NUM_GROUPS);
595 #else // defined(NUM_GROUPS) 596 const int xo = ch * 121;
597 #endif // defined(NUM_GROUPS) 598 const int yo = xc + yc * CONVOLVED_WIDTH;
601 __global uchar *input_ptr = src_ptr + src_offset_first_element_in_bytes + xi * src_stride_x + yi * src_stride_y + ch * src_stride_z + batch * src_stride_w;
602 #if defined(NUM_GROUPS) 603 __global uchar *output_ptr = dst_ptr + dst_offset_first_element_in_bytes + xo * dst_stride_x + yo * dst_stride_y + zo * dst_stride_z + batch * dst_stride_w;
604 #else // defined(NUM_GROUPS) 605 __global uchar *output_ptr = dst_ptr + dst_offset_first_element_in_bytes + xo * dst_stride_x + yo * dst_stride_y + batch * dst_stride_w;
606 #endif // defined(NUM_GROUPS) 610 row00 = vload8(0, (__global DATA_TYPE *)(input_ptr));
612 row01 = vload3(0, (__global DATA_TYPE *)(input_ptr) + 8);
614 vstore8((
VEC_DATA_TYPE(DATA_TYPE, 8))(row00.s01234567), 0, (__global DATA_TYPE *)output_ptr);
615 vstore3((
VEC_DATA_TYPE(DATA_TYPE, 3))(row01.s012), 0, (__global DATA_TYPE *)output_ptr + 8);
617 input_ptr += src_stride_y;
618 output_ptr += 11 * src_stride_x;
623 row00 = vload8(0, (__global DATA_TYPE *)(input_ptr));
625 row01 = vload3(0, (__global DATA_TYPE *)(input_ptr) + 8);
627 vstore8((
VEC_DATA_TYPE(DATA_TYPE, 8))(row00.s01234567), 0, (__global DATA_TYPE *)output_ptr);
628 vstore3((
VEC_DATA_TYPE(DATA_TYPE, 3))(row01.s012), 0, (__global DATA_TYPE *)output_ptr + 8);
630 input_ptr += src_stride_y;
631 output_ptr += 11 * src_stride_x;
636 row00 = vload8(0, (__global DATA_TYPE *)(input_ptr));
638 row01 = vload3(0, (__global DATA_TYPE *)(input_ptr) + 8);
640 vstore8((
VEC_DATA_TYPE(DATA_TYPE, 8))(row00.s01234567), 0, (__global DATA_TYPE *)output_ptr);
641 vstore3((
VEC_DATA_TYPE(DATA_TYPE, 3))(row01.s012), 0, (__global DATA_TYPE *)output_ptr + 8);
643 input_ptr += src_stride_y;
644 output_ptr += 11 * src_stride_x;
649 row00 = vload8(0, (__global DATA_TYPE *)(input_ptr));
651 row01 = vload3(0, (__global DATA_TYPE *)(input_ptr) + 8);
653 vstore8((
VEC_DATA_TYPE(DATA_TYPE, 8))(row00.s01234567), 0, (__global DATA_TYPE *)output_ptr);
654 vstore3((
VEC_DATA_TYPE(DATA_TYPE, 3))(row01.s012), 0, (__global DATA_TYPE *)output_ptr + 8);
656 input_ptr += src_stride_y;
657 output_ptr += 11 * src_stride_x;
662 row00 = vload8(0, (__global DATA_TYPE *)(input_ptr));
664 row01 = vload3(0, (__global DATA_TYPE *)(input_ptr) + 8);
666 vstore8((
VEC_DATA_TYPE(DATA_TYPE, 8))(row00.s01234567), 0, (__global DATA_TYPE *)output_ptr);
667 vstore3((
VEC_DATA_TYPE(DATA_TYPE, 3))(row01.s012), 0, (__global DATA_TYPE *)output_ptr + 8);
669 input_ptr += src_stride_y;
670 output_ptr += 11 * src_stride_x;
675 row00 = vload8(0, (__global DATA_TYPE *)(input_ptr));
677 row01 = vload3(0, (__global DATA_TYPE *)(input_ptr) + 8);
679 vstore8((
VEC_DATA_TYPE(DATA_TYPE, 8))(row00.s01234567), 0, (__global DATA_TYPE *)output_ptr);
680 vstore3((
VEC_DATA_TYPE(DATA_TYPE, 3))(row01.s012), 0, (__global DATA_TYPE *)output_ptr + 8);
682 input_ptr += src_stride_y;
683 output_ptr += 11 * src_stride_x;
688 row00 = vload8(0, (__global DATA_TYPE *)(input_ptr));
690 row01 = vload3(0, (__global DATA_TYPE *)(input_ptr) + 8);
692 vstore8((
VEC_DATA_TYPE(DATA_TYPE, 8))(row00.s01234567), 0, (__global DATA_TYPE *)output_ptr);
693 vstore3((
VEC_DATA_TYPE(DATA_TYPE, 3))(row01.s012), 0, (__global DATA_TYPE *)output_ptr + 8);
695 input_ptr += src_stride_y;
696 output_ptr += 11 * src_stride_x;
701 row00 = vload8(0, (__global DATA_TYPE *)(input_ptr));
703 row01 = vload3(0, (__global DATA_TYPE *)(input_ptr) + 8);
705 vstore8((
VEC_DATA_TYPE(DATA_TYPE, 8))(row00.s01234567), 0, (__global DATA_TYPE *)output_ptr);
706 vstore3((
VEC_DATA_TYPE(DATA_TYPE, 3))(row01.s012), 0, (__global DATA_TYPE *)output_ptr + 8);
708 input_ptr += src_stride_y;
709 output_ptr += 11 * src_stride_x;
714 row00 = vload8(0, (__global DATA_TYPE *)(input_ptr));
716 row01 = vload3(0, (__global DATA_TYPE *)(input_ptr) + 8);
718 vstore8((
VEC_DATA_TYPE(DATA_TYPE, 8))(row00.s01234567), 0, (__global DATA_TYPE *)output_ptr);
719 vstore3((
VEC_DATA_TYPE(DATA_TYPE, 3))(row01.s012), 0, (__global DATA_TYPE *)output_ptr + 8);
721 input_ptr += src_stride_y;
722 output_ptr += 11 * src_stride_x;
727 row00 = vload8(0, (__global DATA_TYPE *)(input_ptr));
729 row01 = vload3(0, (__global DATA_TYPE *)(input_ptr) + 8);
731 vstore8((
VEC_DATA_TYPE(DATA_TYPE, 8))(row00.s01234567), 0, (__global DATA_TYPE *)output_ptr);
732 vstore3((
VEC_DATA_TYPE(DATA_TYPE, 3))(row01.s012), 0, (__global DATA_TYPE *)output_ptr + 8);
734 input_ptr += src_stride_y;
735 output_ptr += 11 * src_stride_x;
740 row00 = vload8(0, (__global DATA_TYPE *)(input_ptr));
742 row01 = vload3(0, (__global DATA_TYPE *)(input_ptr) + 8);
744 vstore8((
VEC_DATA_TYPE(DATA_TYPE, 8))(row00.s01234567), 0, (__global DATA_TYPE *)output_ptr);
745 vstore3((
VEC_DATA_TYPE(DATA_TYPE, 3))(row01.s012), 0, (__global DATA_TYPE *)output_ptr + 8);
747 output_ptr += 11 * src_stride_x;
751 #if defined(NUM_GROUPS) 752 if((xo / 121) == (SRC_DEPTH / NUM_GROUPS - 1))
753 #else // defined(NUM_GROUPS) 754 if(ch == (SRC_DEPTH - 1))
755 #endif // defined(NUM_GROUPS) 757 *((__global DATA_TYPE *)output_ptr) = 1.0f;
761 #endif // defined(CONVOLVED_WIDTH) && defined(STRIDE_X) && defined(STRIDE_Y) && defined(SRC_DEPTH) 763 #if defined(CONVOLVED_WIDTH) && defined(STRIDE_X) && defined(STRIDE_Y) && defined(KERNEL_WIDTH) && defined(KERNEL_HEIGHT) && defined(SRC_DEPTH) && defined(SRC_WIDTH) && defined(SRC_HEIGHT) && defined(VECTOR_SIZE) && defined(WIDTH_MOD_VECTOR_SIZE) 792 __kernel
void im2col_generic_padx0_pady0_nchw(
794 #
if defined(NUM_GROUPS)
802 const int xc = get_global_id(0);
803 const int yc = get_global_id(1);
804 const int ch = get_global_id(2) % SRC_DEPTH;
805 const int batch = get_global_id(2) / SRC_DEPTH;
808 const int xi = xc * STRIDE_X;
809 const int yi = yc * STRIDE_Y;
812 #if defined(NUM_GROUPS) 813 const int xo = (ch % (SRC_DEPTH / NUM_GROUPS)) * KERNEL_WIDTH * KERNEL_HEIGHT;
814 const int zo = ch / (SRC_DEPTH / NUM_GROUPS);
815 #else // defined(NUM_GROUPS) 816 const int xo = ch * KERNEL_WIDTH * KERNEL_HEIGHT;
817 #endif // defined(NUM_GROUPS) 818 const int yo = xc + yc * CONVOLVED_WIDTH;
820 __global uchar *input_ptr = src_ptr + src_offset_first_element_in_bytes + ch * src_stride_z + batch * src_stride_w;
821 #if defined(NUM_GROUPS) 822 __global DATA_TYPE *output_ptr = ((__global DATA_TYPE *)(dst_ptr + dst_offset_first_element_in_bytes + yo * dst_stride_y + zo * dst_stride_z + batch * dst_stride_w)) + xo;
823 #else // defined(NUM_GROUPS) 824 __global DATA_TYPE *output_ptr = ((__global DATA_TYPE *)(dst_ptr + dst_offset_first_element_in_bytes + yo * dst_stride_y + batch * dst_stride_w)) + xo;
825 #endif // defined(NUM_GROUPS) 828 for(
int y = yi, y_e = yi + KERNEL_HEIGHT; y < y_e; ++y)
831 for(
int x = xi, x_e = xi + KERNEL_WIDTH; x + VECTOR_SIZE <= x_e; x += VECTOR_SIZE, output_ptr += VECTOR_SIZE)
834 row =
VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + x * src_stride_x + y * src_stride_y));
836 (row, 0, output_ptr);
841 #if WIDTH_MOD_VECTOR_SIZE == 1 842 *output_ptr = *((__global DATA_TYPE *)(input_ptr + (last_x + VECTOR_SIZE) * src_stride_x + y * src_stride_y));
843 #elif WIDTH_MOD_VECTOR_SIZE > 1 845 row =
VLOAD(WIDTH_MOD_VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + (last_x + VECTOR_SIZE) * src_stride_x + y * src_stride_y));
846 VSTORE(WIDTH_MOD_VECTOR_SIZE)
847 (row, 0, output_ptr);
849 output_ptr += WIDTH_MOD_VECTOR_SIZE;
853 #if defined(NUM_GROUPS) 854 if((xo / (KERNEL_WIDTH * KERNEL_HEIGHT)) == (SRC_DEPTH / NUM_GROUPS - 1))
855 #else // defined(NUM_GROUPS) 856 if(ch == (SRC_DEPTH - 1))
863 #endif //defined(CONVOLVED_WIDTH) && defined(STRIDE_X) && defined(STRIDE_Y) && defined(PAD_LEFT) && defined(PAD_TOP) && defined(PAD_RIGHT) && defined(PAD_BOTTOM) && defined(KERNEL_WIDTH) && defined(KERNEL_HEIGHT) && defined(SRC_DEPTH) && defined(SRC_WIDTH) && defined(SRC_HEIGHT) && defined(VECTOR_SIZE) && defined(WIDTH_MOD_VECTOR_SIZE) 865 #if defined(CONVOLVED_WIDTH) && defined(SRC_WIDTH) && defined(SRC_HEIGHT) && defined(STRIDE_X) && defined(STRIDE_Y) && defined(KERNEL_WIDTH) && defined(KERNEL_HEIGHT) && defined(SRC_DEPTH) && defined(PAD_LEFT) && defined(PAD_RIGHT) && defined(PAD_TOP) && defined(PAD_BOTTOM) && defined(PAD_VALUE) && defined(VECTOR_SIZE) && defined(BOUNDARY_VECTOR_SIZE) 867 #define VECTOR_N VEC_DATA_TYPE(DATA_TYPE, VECTOR_SIZE) 868 #define COND_N VEC_DATA_TYPE(COND_DATA_TYPE, VECTOR_SIZE) 884 #if defined(VECTOR_SIZE) && defined(BOUNDARY_VECTOR_SIZE) && BOUNDARY_VECTOR_SIZE < VECTOR_SIZE 885 #define IM2COL1X9_NHWC_STORE(VECTOR_SIZE, BOUNDARY_VECTOR_SIZE, DATA_TYPE, SRC_DEPTH, DATA, ROW, OUTPUT_PTR) \ 886 const bool at_channel_boundary = get_global_id(0) == 0; \ 887 if(at_channel_boundary) \ 889 IM2COL1X9_NHWC_STORE_PARTIAL(VECTOR_SIZE, BOUNDARY_VECTOR_SIZE, DATA_TYPE, SRC_DEPTH, DATA, ROW, OUTPUT_PTR) \ 893 IM2COL1X9_NHWC_STORE_NONPARTIAL(VECTOR_SIZE, DATA_TYPE, SRC_DEPTH, DATA, ROW, OUTPUT_PTR) \ 895 #else // defined(VECTOR_SIZE) && defined(BOUNDARY_VECTOR_SIZE) && BOUNDARY_VECTOR_SIZE < VECTOR_SIZE 896 #define IM2COL1X9_NHWC_STORE(VECTOR_SIZE, BOUNDARY_VECTOR_SIZE, DATA_TYPE, SRC_DEPTH, DATA, ROW, OUTPUT_PTR) \ 897 IM2COL1X9_NHWC_STORE_NONPARTIAL(VECTOR_SIZE, DATA_TYPE, SRC_DEPTH, DATA, ROW, OUTPUT_PTR) 898 #endif // defined(VECTOR_SIZE) && defined(BOUNDARY_VECTOR_SIZE) && BOUNDARY_VECTOR_SIZE < VECTOR_SIZE 900 #define IM2COL1X9_NHWC_STORE_NONPARTIAL(VECTOR_SIZE, DATA_TYPE, SRC_DEPTH, DATA, ROW, OUTPUT_PTR) \ 901 VSTORE(VECTOR_SIZE) \ 902 (DATA##0, 0, (__global DATA_TYPE *)(OUTPUT_PTR) + (0 + ROW * 9) * SRC_DEPTH); \ 903 VSTORE(VECTOR_SIZE) \ 904 (DATA##1, 0, (__global DATA_TYPE *)(OUTPUT_PTR) + (1 + ROW * 9) * SRC_DEPTH); \ 905 VSTORE(VECTOR_SIZE) \ 906 (DATA##2, 0, (__global DATA_TYPE *)(OUTPUT_PTR) + (2 + ROW * 9) * SRC_DEPTH); \ 907 VSTORE(VECTOR_SIZE) \ 908 (DATA##3, 0, (__global DATA_TYPE *)(OUTPUT_PTR) + (3 + ROW * 9) * SRC_DEPTH); \ 909 VSTORE(VECTOR_SIZE) \ 910 (DATA##4, 0, (__global DATA_TYPE *)(OUTPUT_PTR) + (4 + ROW * 9) * SRC_DEPTH); \ 911 VSTORE(VECTOR_SIZE) \ 912 (DATA##5, 0, (__global DATA_TYPE *)(OUTPUT_PTR) + (5 + ROW * 9) * SRC_DEPTH); \ 913 VSTORE(VECTOR_SIZE) \ 914 (DATA##6, 0, (__global DATA_TYPE *)(OUTPUT_PTR) + (6 + ROW * 9) * SRC_DEPTH); \ 915 VSTORE(VECTOR_SIZE) \ 916 (DATA##7, 0, (__global DATA_TYPE *)(OUTPUT_PTR) + (7 + ROW * 9) * SRC_DEPTH); \ 917 VSTORE(VECTOR_SIZE) \ 918 (DATA##8, 0, (__global DATA_TYPE *)(OUTPUT_PTR) + (8 + ROW * 9) * SRC_DEPTH); 920 #define IM2COL1X9_NHWC_STORE_PARTIAL(VECTOR_SIZE, BOUNDARY_VECTOR_SIZE, DATA_TYPE, SRC_DEPTH, DATA, ROW, OUTPUT_PTR) \ 921 VSTORE_PARTIAL(VECTOR_SIZE, BOUNDARY_VECTOR_SIZE) \ 922 (DATA##0, 0, (__global DATA_TYPE *)(OUTPUT_PTR) + (0 + ROW * 9) * SRC_DEPTH); \ 923 VSTORE_PARTIAL(VECTOR_SIZE, BOUNDARY_VECTOR_SIZE) \ 924 (DATA##1, 0, (__global DATA_TYPE *)(OUTPUT_PTR) + (1 + ROW * 9) * SRC_DEPTH); \ 925 VSTORE_PARTIAL(VECTOR_SIZE, BOUNDARY_VECTOR_SIZE) \ 926 (DATA##2, 0, (__global DATA_TYPE *)(OUTPUT_PTR) + (2 + ROW * 9) * SRC_DEPTH); \ 927 VSTORE_PARTIAL(VECTOR_SIZE, BOUNDARY_VECTOR_SIZE) \ 928 (DATA##3, 0, (__global DATA_TYPE *)(OUTPUT_PTR) + (3 + ROW * 9) * SRC_DEPTH); \ 929 VSTORE_PARTIAL(VECTOR_SIZE, BOUNDARY_VECTOR_SIZE) \ 930 (DATA##4, 0, (__global DATA_TYPE *)(OUTPUT_PTR) + (4 + ROW * 9) * SRC_DEPTH); \ 931 VSTORE_PARTIAL(VECTOR_SIZE, BOUNDARY_VECTOR_SIZE) \ 932 (DATA##5, 0, (__global DATA_TYPE *)(OUTPUT_PTR) + (5 + ROW * 9) * SRC_DEPTH); \ 933 VSTORE_PARTIAL(VECTOR_SIZE, BOUNDARY_VECTOR_SIZE) \ 934 (DATA##6, 0, (__global DATA_TYPE *)(OUTPUT_PTR) + (6 + ROW * 9) * SRC_DEPTH); \ 935 VSTORE_PARTIAL(VECTOR_SIZE, BOUNDARY_VECTOR_SIZE) \ 936 (DATA##7, 0, (__global DATA_TYPE *)(OUTPUT_PTR) + (7 + ROW * 9) * SRC_DEPTH); \ 937 VSTORE_PARTIAL(VECTOR_SIZE, BOUNDARY_VECTOR_SIZE) \ 938 (DATA##8, 0, (__global DATA_TYPE *)(OUTPUT_PTR) + (8 + ROW * 9) * SRC_DEPTH); 970 __kernel
void im2col3x3_nhwc(
977 const int shift_amount = (int)VECTOR_SIZE - (
int)BOUNDARY_VECTOR_SIZE;
978 const int ch = max((
int)(get_global_id(0) * VECTOR_SIZE) - shift_amount, 0);
979 const int yo = get_global_id(1);
980 const int batch = get_global_id(2);
983 const int xi = (get_global_id(1) % CONVOLVED_WIDTH) * STRIDE_X;
984 const int yi = (get_global_id(1) / (int)CONVOLVED_WIDTH) * STRIDE_Y;
987 __global uchar *input_ptr = src_ptr + src_offset_first_element_in_bytes + ch *
sizeof(
DATA_TYPE) + batch * (
int)src_stride_w;
988 __global uchar *output_ptr = dst_ptr + dst_offset_first_element_in_bytes + ch *
sizeof(
DATA_TYPE) + yo * (
int)dst_stride_y + batch * (int)dst_stride_w;
994 int3 xi_offset = ((int3)xi + (int3)(0, 1, 2) * DILATION_X - (int3)PAD_LEFT);
995 #if PAD_LEFT != 0 || PAD_RIGHT != 0 996 #define CLAMP(x, min_val, max_val) min(max(x, min_val), max_val) 997 xi_offset =
CLAMP(xi_offset, (int3)0, (int3)(SRC_WIDTH - 1));
998 #endif // PAD_LEFT != 0 || PAD_RIGHT != 0 1000 xi_offset *= (int3)src_stride_y;
1003 int3 x_cond = (((int3)xi + (int3)(0, 1, 2) * DILATION_X - (int3)PAD_LEFT) < (int3)0) || (((int3)xi + (int3)(0, 1, 2) * DILATION_X - (int3)PAD_LEFT) >= (int3)SRC_WIDTH);
1012 yi_coord = yi - (int)PAD_TOP;
1015 #if PAD_TOP != 0 || PAD_BOTTOM != 0 1016 yi_coord = min((uint)yi_coord, (uint)(SRC_HEIGHT - 1));
1017 #endif // PAD_TOP != 0 || PAD_BOTTOM != 0 1020 offset = xi_offset + (yi_coord * (int)src_stride_z);
1023 VECTOR_N values0 =
VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset.s0));
1024 VECTOR_N values1 =
VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset.s1));
1025 VECTOR_N values2 =
VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset.s2));
1027 #if PAD_TOP != 0 || PAD_LEFT != 0 || PAD_BOTTOM != 0 || PAD_RIGHT != 0 1029 int y_cond = (int)((uint)(yi - (int)PAD_TOP) >= (uint)(SRC_HEIGHT));
1030 values0 =
select(values0, (VECTOR_N)PAD_VALUE, (COND_N)((COND_N)y_cond || (COND_N)(x_cond.s0)));
1031 values1 =
select(values1, (VECTOR_N)PAD_VALUE, (COND_N)((COND_N)y_cond || (COND_N)(x_cond.s1)));
1032 values2 =
select(values2, (VECTOR_N)PAD_VALUE, (COND_N)((COND_N)y_cond || (COND_N)(x_cond.s2)));
1033 #endif // PAD_TOP != 0 || PAD_LEFT != 0 || PAD_BOTTOM != 0 || PAD_RIGHT != 0 1037 yi_coord = yi - (int)PAD_TOP + 1 * DILATION_Y;
1040 #if PAD_TOP != 0 || PAD_BOTTOM != 0 1041 yi_coord = min((uint)yi_coord, (uint)(SRC_HEIGHT - 1));
1042 #endif // PAD_TOP != 0 || PAD_BOTTOM != 0 1045 offset = xi_offset + (yi_coord * (int)src_stride_z);
1048 VECTOR_N values3 =
VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset.s0));
1049 VECTOR_N values4 =
VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset.s1));
1050 VECTOR_N values5 =
VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset.s2));
1052 #if PAD_TOP != 0 || PAD_LEFT != 0 || PAD_BOTTOM != 0 || PAD_RIGHT != 0 1054 y_cond = (int)((uint)(yi - (int)PAD_TOP + 1 * DILATION_Y) >= (uint)(SRC_HEIGHT));
1055 values3 =
select(values3, (VECTOR_N)PAD_VALUE, (COND_N)((COND_N)y_cond || (COND_N)(x_cond.s0)));
1056 values4 =
select(values4, (VECTOR_N)PAD_VALUE, (COND_N)((COND_N)y_cond || (COND_N)(x_cond.s1)));
1057 values5 =
select(values5, (VECTOR_N)PAD_VALUE, (COND_N)((COND_N)y_cond || (COND_N)(x_cond.s2)));
1058 #endif // PAD_TOP != 0 || PAD_LEFT != 0 || PAD_BOTTOM != 0 || PAD_RIGHT != 0 1062 yi_coord = yi - (int)PAD_TOP + 2 * DILATION_Y;
1065 #if PAD_TOP != 0 || PAD_BOTTOM != 0 1066 yi_coord = min((uint)yi_coord, (uint)(SRC_HEIGHT - 1));
1067 #endif // PAD_TOP != 0 || PAD_BOTTOM != 0 1070 offset = xi_offset + (yi_coord * (int)src_stride_z);
1073 VECTOR_N values6 =
VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset.s0));
1074 VECTOR_N values7 =
VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset.s1));
1075 VECTOR_N values8 =
VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset.s2));
1077 #if PAD_TOP != 0 || PAD_LEFT != 0 || PAD_BOTTOM != 0 || PAD_RIGHT != 0 1079 y_cond = (int)((uint)(yi - (int)PAD_TOP + 2 * DILATION_Y) >= (uint)(SRC_HEIGHT));
1080 values6 =
select(values6, (VECTOR_N)PAD_VALUE, (COND_N)((COND_N)y_cond || (COND_N)(x_cond.s0)));
1081 values7 =
select(values7, (VECTOR_N)PAD_VALUE, (COND_N)((COND_N)y_cond || (COND_N)(x_cond.s1)));
1082 values8 =
select(values8, (VECTOR_N)PAD_VALUE, (COND_N)((COND_N)y_cond || (COND_N)(x_cond.s2)));
1083 #endif // PAD_TOP != 0 || PAD_LEFT != 0 || PAD_BOTTOM != 0 || PAD_RIGHT != 0 1086 IM2COL1X9_NHWC_STORE(VECTOR_SIZE, BOUNDARY_VECTOR_SIZE, DATA_TYPE, SRC_DEPTH, values, 0, output_ptr)
1094 if((ch + VECTOR_SIZE) >= SRC_DEPTH)
1096 *((__global DATA_TYPE *)(output_ptr) - ch + SRC_DEPTH * 9) = 1.0f;
1101 #if PAD_TOP != 0 || PAD_LEFT != 0 || PAD_BOTTOM != 0 || PAD_RIGHT != 0 1102 #define IM2COL1x9(i) \ 1104 yi_coord = yi - (int)PAD_TOP + i * DILATION_Y; \ 1105 yi_coord = min((uint)yi_coord, (uint)(SRC_HEIGHT - 1)); \ 1107 offset0 = xi_offset0 + (yi_coord * (int)src_stride_z); \ 1108 offset1 = xi_offset1 + (yi_coord * (int)src_stride_z); \ 1110 VECTOR_N values0 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset0.s0)); \ 1111 VECTOR_N values1 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset0.s1)); \ 1112 VECTOR_N values2 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset0.s2)); \ 1113 VECTOR_N values3 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset0.s3)); \ 1114 VECTOR_N values4 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset0.s4)); \ 1115 VECTOR_N values5 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset0.s5)); \ 1116 VECTOR_N values6 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset0.s6)); \ 1117 VECTOR_N values7 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset0.s7)); \ 1118 VECTOR_N values8 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset1)); \ 1120 int y_cond = (int)((uint)(yi - (int)PAD_TOP + i * DILATION_Y) >= (uint)(SRC_HEIGHT)); \ 1121 values0 = select(values0, (VECTOR_N)PAD_VALUE, (COND_N)((COND_N)y_cond || (COND_N)(x_cond0.s0))); \ 1122 values1 = select(values1, (VECTOR_N)PAD_VALUE, (COND_N)((COND_N)y_cond || (COND_N)(x_cond0.s1))); \ 1123 values2 = select(values2, (VECTOR_N)PAD_VALUE, (COND_N)((COND_N)y_cond || (COND_N)(x_cond0.s2))); \ 1124 values3 = select(values3, (VECTOR_N)PAD_VALUE, (COND_N)((COND_N)y_cond || (COND_N)(x_cond0.s3))); \ 1125 values4 = select(values4, (VECTOR_N)PAD_VALUE, (COND_N)((COND_N)y_cond || (COND_N)(x_cond0.s4))); \ 1126 values5 = select(values5, (VECTOR_N)PAD_VALUE, (COND_N)((COND_N)y_cond || (COND_N)(x_cond0.s5))); \ 1127 values6 = select(values6, (VECTOR_N)PAD_VALUE, (COND_N)((COND_N)y_cond || (COND_N)(x_cond0.s6))); \ 1128 values7 = select(values7, (VECTOR_N)PAD_VALUE, (COND_N)((COND_N)y_cond || (COND_N)(x_cond0.s7))); \ 1129 values8 = select(values8, (VECTOR_N)PAD_VALUE, (COND_N)((COND_N)y_cond || (COND_N)(x_cond1))); \ 1131 IM2COL1X9_NHWC_STORE(VECTOR_SIZE, BOUNDARY_VECTOR_SIZE, DATA_TYPE, SRC_DEPTH, values, i, output_ptr) \ 1133 #else // PAD_TOP != 0 || PAD_LEFT != 0 || PAD_BOTTOM != 0 || PAD_RIGHT != 0 1134 #define IM2COL1x9(i) \ 1136 yi_coord = yi - (int)PAD_TOP + i * DILATION_Y; \ 1137 yi_coord = min((uint)yi_coord, (uint)(SRC_HEIGHT - 1)); \ 1139 offset0 = xi_offset0 + (yi_coord * (int)src_stride_z); \ 1140 offset1 = xi_offset1 + (yi_coord * (int)src_stride_z); \ 1142 VECTOR_N values0 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset0.s0)); \ 1143 VECTOR_N values1 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset0.s1)); \ 1144 VECTOR_N values2 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset0.s2)); \ 1145 VECTOR_N values3 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset0.s3)); \ 1146 VECTOR_N values4 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset0.s4)); \ 1147 VECTOR_N values5 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset0.s5)); \ 1148 VECTOR_N values6 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset0.s6)); \ 1149 VECTOR_N values7 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset0.s7)); \ 1150 VECTOR_N values8 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset1)); \ 1152 IM2COL1X9_NHWC_STORE(VECTOR_SIZE, BOUNDARY_VECTOR_SIZE, DATA_TYPE, SRC_DEPTH, values, i, output_ptr) \ 1154 #endif // PAD_TOP != 0 || PAD_LEFT != 0 || PAD_BOTTOM != 0 || PAD_RIGHT != 0 1185 __kernel
void im2col9x9_nhwc(
1192 const int shift_amount = (int)VECTOR_SIZE - (
int)BOUNDARY_VECTOR_SIZE;
1193 const int ch = max((
int)(get_global_id(0) * VECTOR_SIZE) - shift_amount, 0);
1194 const int yo = get_global_id(1);
1195 const int batch = get_global_id(2);
1198 const int xi = (get_global_id(1) % CONVOLVED_WIDTH) * STRIDE_X;
1199 const int yi = (get_global_id(1) / (int)CONVOLVED_WIDTH) * STRIDE_Y;
1202 __global uchar *input_ptr = src_ptr + src_offset_first_element_in_bytes + ch *
sizeof(
DATA_TYPE) + batch * (
int)src_stride_w;
1203 __global uchar *output_ptr = dst_ptr + dst_offset_first_element_in_bytes + ch *
sizeof(
DATA_TYPE) + yo * (
int)dst_stride_y + batch * (int)dst_stride_w;
1210 int8 xi_offset0 = ((int8)xi + (int8)(0, 1, 2, 3, 4, 5, 6, 7) * DILATION_X - (int8)PAD_LEFT);
1211 int xi_offset1 = ((int)xi + (
int)(8) * DILATION_X - (
int)PAD_LEFT);
1213 #if PAD_LEFT != 0 || PAD_RIGHT != 0 1214 #define CLAMP(x, min_val, max_val) min(max(x, min_val), max_val) 1215 xi_offset0 =
CLAMP(xi_offset0, (int8)0, (int8)(SRC_WIDTH - 1));
1216 xi_offset1 =
CLAMP(xi_offset1, (
int)0, (
int)(SRC_WIDTH - 1));
1217 #endif // PAD_LEFT != 0 || PAD_RIGHT != 0 1218 xi_offset0 *= (int8)src_stride_y;
1219 xi_offset1 *= (int)src_stride_y;
1222 int8 x_cond0 = (((int8)xi + (int8)(0, 1, 2, 3, 4, 5, 6, 7) * DILATION_X - (int8)PAD_LEFT) < (int8)0) || (((int8)xi + (int8)(0, 1, 2, 3, 4, 5, 6, 7) * DILATION_X - (int8)PAD_LEFT) >= (int8)SRC_WIDTH);
1223 int x_cond1 = (((int)xi + (
int)(8) * DILATION_X - (
int)PAD_LEFT) < (
int)0) || (((
int)xi + (int)(8) * DILATION_X - (int)PAD_LEFT) >= (int)SRC_WIDTH);
1241 if((ch + VECTOR_SIZE) >= SRC_DEPTH)
1243 *((__global DATA_TYPE *)(output_ptr) - ch + SRC_DEPTH * 81) = 1.0f;
1281 __kernel
void im2col_generic_nhwc(
1288 const int shift_amount = (int)VECTOR_SIZE - (
int)BOUNDARY_VECTOR_SIZE;
1289 const int ch = max((
int)(get_global_id(0) * VECTOR_SIZE) - shift_amount, 0);
1290 const int yo = get_global_id(1);
1291 const int batch = get_global_id(2);
1294 const int xi = (get_global_id(1) % CONVOLVED_WIDTH) * STRIDE_X;
1295 const int yi = (get_global_id(1) / (int)CONVOLVED_WIDTH) * STRIDE_Y;
1298 __global uchar *input_ptr = src_ptr + src_offset_first_element_in_bytes + ch *
sizeof(
DATA_TYPE) + batch * (
int)src_stride_w;
1299 __global uchar *output_ptr = dst_ptr + dst_offset_first_element_in_bytes + ch *
sizeof(
DATA_TYPE) + yo * (
int)dst_stride_y + batch * (int)dst_stride_w;
1302 for(
int yk = 0; yk < KERNEL_HEIGHT; ++yk)
1305 int yi_coord = yi + yk * DILATION_Y - (int)PAD_TOP;
1306 yi_coord =
CLAMP(yi_coord, (
int)0, (
int)(SRC_HEIGHT - 1));
1309 int y_border_condition = ((yi + yk * DILATION_Y - (int)PAD_TOP) < (int)0) || ((yi + yk * DILATION_Y - (int)PAD_TOP) >= (int)SRC_HEIGHT);
1311 for(
int xk = 0; xk < KERNEL_WIDTH; ++xk)
1314 int xi_coord = (xi + xk * DILATION_X - (int)PAD_LEFT);
1315 xi_coord =
CLAMP(xi_coord, (
int)0, (
int)(SRC_WIDTH - 1));
1318 int x_border_condition = ((xi + xk * DILATION_X - (int)PAD_LEFT) < (int)0) || ((xi + xk * DILATION_X - (int)PAD_LEFT) >= (int)SRC_WIDTH);
1320 int offset = xi_coord * (int)src_stride_y + (yi_coord * (
int)src_stride_z);
1322 VECTOR_N values0 =
VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset));
1324 #if PAD_LEFT != 0 || PAD_TOP != 0 || PAD_RIGHT != 0 || PAD_BOTTOM != 0 1326 values0 =
select(values0, (VECTOR_N)PAD_VALUE, (COND_N)((COND_N)x_border_condition || (COND_N)(y_border_condition)));
1327 #endif // PAD_LEFT != 0 || PAD_TOP != 0 || PAD_RIGHT != 0 || PAD_BOTTOM != 0 1330 #if BOUNDARY_VECTOR_SIZE != VECTOR_SIZE 1331 const bool at_channel_boundary = get_global_id(0) == 0;
1332 if(at_channel_boundary)
1335 (values0, 0, (__global DATA_TYPE *)(output_ptr) + i * (int)SRC_DEPTH);
1338 #endif // BOUNDARY_VECTOR_SIZE != VECTOR_SIZE 1341 (values0, 0, (__global DATA_TYPE *)(output_ptr) + i * (int)SRC_DEPTH);
1353 if((ch + VECTOR_SIZE) >= SRC_DEPTH)
1355 *((__global DATA_TYPE *)(output_ptr) - ch + SRC_DEPTH * KERNEL_WIDTH * KERNEL_HEIGHT) = 1.0f;
1359 #endif // defined(CONVOLVED_WIDTH) && defined(SRC_WIDTH) && defined(SRC_HEIGHT) && defined(STRIDE_X) && defined(STRIDE_Y) && defined(KERNEL_WIDTH) && defined(KERNEL_HEIGHT) && defined(SRC_DEPTH) && defined(PAD_LEFT) && defined(PAD_RIGHT) && defined(PAD_TOP) && defined(PAD_BOTTOM) && defined(PAD_VALUE) && defined(VECTOR_SIZE) && defined(BOUNDARY_VECTOR_SIZE) 1360 #endif // defined(DATA_TYPE) && defined(ELEMENT_SIZE) __global uchar * offset(const Image *img, int x, int y)
Get the pointer position of a Image.
#define VSTORE_PARTIAL(size, store_size)
#define IMAGE_DECLARATION(name)
SimpleTensor< float > src
#define CLAMP(x, min_val, max_val)
Clamp the given value between an upper and lower bound.
SimpleTensor< T > select(const SimpleTensor< uint8_t > &c, const SimpleTensor< T > &x, const SimpleTensor< T > &y)
#define TENSOR3D_DECLARATION(name)
#define VEC_DATA_TYPE(type, size)